LeRobot

In June of 2025 the Brain Wave Collective hosted Denver for HuggingFace's Global Hackathon

HuggingFace's LeRobot is an impressive Python based open-source project by which emphasises accessable introduction to reproducible datasets, training model policies, and community-driven tooling. There is perhaps no better entry point for exploring learning-from-demonstrations workflows and the collection of datasets, as well as training and evaluation of policy. The project serves as a base framework for a number of our derivative builds such as Von Puppenstein, Trash Detector, LeRobo-Vous, and Yes Chef!.

The project enables low-cost exploration of novel AI/ML robotics approaches, such as ACT (A novel approach to fine-grained bimanual manipulation). The project has also gained broader recognition and industry integration such as when NVIDIA released the impressive GR00T N1.5 and only a few weeks later published a guide for post-training N1.5 with the LeRobot SO-101 arm. Additional models are available with many other supported uses.

Although LeRobot can be adopted to a wide range of robotic arms, the most prominently featured is the 3D printable SO-101 designed by Rob Knight from The Robot Studio.

Another interesting development to come out of the project was when HuggingFace released SmolVLM, an impressively small visual language to action model capable of running on commodity hardware. What made this interesting is that it that this achieved SOTA result while be trained entirely on public datasets produced by the LeRobot project.